I do a lot of work in similar areas here.

While vibe coding will get you something that potentially works, I've noticed LLMs are really bad at cleanly abstracting across multiple layers in this area. They usually will insist on parsing and serializing every field at every layer.

If you have the protocols/interfaces well defined up front it is very fast at building extensions, analytics or visualizations though.

> I've noticed LLMs are really bad at cleanly abstracting across multiple layers

Which makes sense, as most developers are too (it’s a particular non-trivial skill and rarely modeled wrll), so LLMs are more likely to be trained on muddled multiple layers.